The 212AY Library
AI
Glossary.
200+ artificial intelligence terms explained in plain language — from machine learning and LLMs to AI agents, RAG and prompt engineering. Every definition comes with a concrete example, in English, French and Arabic.
200 terms
Evaluation Metrics
Evaluation metrics are quantitative measures assessing how well an AI model performs. Different tasks require different metrics: accuracy and F1 for classification, mean squared error for regression, BLEU for translation, ROUGE for summarization. Choosing the right metric is critical — optimizing for the wrong one can produce a model that scores well on paper but fails in practice. Metrics should reflect the actual business objective.
Explainable AI (XAI)
Explainable AI encompasses techniques that make AI model decisions understandable to humans, addressing the black-box nature of deep neural networks. Methods include SHAP values, LIME, attention visualization, and concept-based explanations. XAI is essential for regulated industries: a Moroccan bank must be able to explain why a loan was denied, and a hospital must justify a diagnostic AI's recommendation to a physician.
Feature Engineering
Feature engineering uses domain knowledge to create, transform, or select input variables that make ML models work more effectively. Examples include converting timestamps into day-of-week features, combining address fields into distance metrics, or creating financial ratios. While deep learning learns features automatically, feature engineering remains essential for tabular business data and often delivers bigger gains than model tuning.
Feature Extraction
Feature extraction is the process of transforming raw data into a set of informative, non-redundant numerical features that machine learning models can use. In computer vision, a CNN automatically extracts edges and textures; in NLP, tokenizers convert words to IDs. Handcrafted features were common before deep learning; now models learn features automatically, though feature engineering remains valuable for tabular business data like customer demographics and transaction histories.
Feature Selection
Feature selection identifies and keeps only the most relevant input variables while discarding redundant or irrelevant ones. Fewer features mean faster training, lower overfitting risk, and easier interpretation. Techniques include correlation analysis, mutual information, recursive elimination, and L1 regularization. A credit scoring model might start with 200 customer attributes and find only 30 drive predictions.
Federated Learning
Federated learning is a technique where multiple organizations collaboratively train an AI model without sharing their raw data. Each participant trains a local copy of the model on their own data, then shares only the model updates — not the data itself — with a central server that aggregates them. It is like several hospitals each studying their own patient records and then pooling what they learned, without any patient file ever leaving the hospital. Privacy and regulatory compliance are the main drivers.
Few-Shot Learning
Few-shot learning is a model's ability to perform a new task after seeing only a handful of examples, typically two to five, without retraining. Large language models excel at this: show the model a few input-output pairs and it infers the pattern. An HR team can classify CVs as junior or senior by providing three labeled examples in the prompt. Few-shot capability is what makes modern LLMs flexible enough for thousands of business use cases without building custom models.
Few-Shot Prompting
Few-shot prompting provides a language model with 2–5 examples of desired input-output behavior directly in the prompt before asking it to perform the task. The model infers the pattern without fine-tuning. For example, listing three customer emails labeled as positive, negative, or neutral, then asking the model to classify a fourth. It is the most practical prompting method for business applications.
Fine-Tuning
Fine-tuning takes a pre-trained model and further trains it on a smaller, task-specific dataset to adapt it to a particular use case. Instead of building from scratch, you start with a model that already understands language and specialize it. A Moroccan bank might fine-tune an open-source LLM on its internal loan policies. Methods range from full fine-tuning to parameter-efficient approaches like LoRA and QLoRA.